The following bug has been logged on the website:
Bug reference: 17478
Logged by: Stefan Pastrilov
Email address: spastrilov@gmail.com
PostgreSQL version: 14.1
Operating system: Ubuntu 20.04
Description:
Hello,
We’re observing a strange bug, according to me. We’ve got native logical
replication where the master host is running on PostgreSQL 13.3 and the
logical replica host is running on PostgreSQL 14.1. And the problem is:
when we start to reindex concurrently the primary key of an 800M heavily
inserted/updated(on the master)table on the logical replica database,
several documents that have been inserted and existing in the master
database during the index rebuild on the replica, are missing from the index
structure on the replica host. What do I mean? If you try to search them by
where id = 5 the index returns no records, but when you force a sequential
scan (where id + 0 = 5)the row is returned. Usually, we’re messing around
10-100 documents from the index after each concurrent re-index. When we
rebuild the index again, they are found by an index look-up, but a new set
of ids are missing. So the problem is easily reproducible, but you should
have a heavily loaded table. Do you have any ideas about this? The newly
created index is valid and fully functional, with no signs of a corrupted
one…
To be clear, the index rebuild is done on the subscribed table(slave
host).
master host:
PostgreSQL 13.3 (Ubuntu 13.3-1.pgdg20.04+1) on x86_64-pc-linux-gnu, compiled
by gcc (Ubuntu 9.3.0-17ubuntu1~20.04) 9.3.0, 64-bit
slave host:
PostgreSQL 14.1 (Ubuntu 14.1-2.pgdg20.04+1) on x86_64-pc-linux-gnu, compiled
by gcc (Ubuntu 9.3.0-17ubuntu1~20.04) 9.3.0, 64-bit